Neural Network based pulse shape analysis with the Belle ...

Neural Network based pulse shape analysis with the Belle II electromagnetic calorimeter5th of November 2020

Women Physicists' Conference

Stella Wermuth

[email protected]

Belle II overview

• SuperKEKB 𝑒+𝑒− collider in Japan

• Collision energy: 10.58 GeV

• Main goals of Belle II:• Precision measurements

• Rare/forbidden processes

• Beyond Standard Model search

https://www.belle2.org/project/super_kekb_and_belle_ii

𝑒−

𝑒+

Stella Wermuth - Belle II Pulse Shape Analysis 2

• ~9000 scintillator crystals CsI(Tl)

• Main task: • Measuring electromagnetic energy

• Reconstructing neutral particles

• Particle identification using shower shapes

Electromagnetic Calorimeter (ECL)

https://www.belle2.org/project/super_kekb_and_belle_ii


Analysis: 𝑒+𝑒− → 𝜋+𝜋−𝛾

Need good particle identification!

Belle II, https://docs.belle2.org/record/1124/files/BELLE2-NOTE-PL-2018-032.pdf


Analysis: 𝑒+𝑒− → 𝜋+𝜋−𝛾Belle II, https://docs.belle2.org/record/1124/files/BELLE2-NOTE-PL-2018-032.pdf A. Keshavarzi, D. Nomura, T. Teubner arXiv:1911.00367 [hep-ph]

Theory Experiment

Improve precision of 𝑒+𝑒− → 𝜋+𝜋−

measurement

Reduce uncertainty of SM-calculation of the anomalous

magnetic momentum of the muon


What happens in the ECL?

AD

C c

ou

nts

digitized pulse shape

6

time

= ECL crystal

= ECL crystal where energy is deposited


What happens in the ECL?

AD

C c

ou

nts

digitized pulse shape

7

time

FPGA• Fit pulse shape• Reconstruct Etotal

• Optimized for photons

FPGA Fit


Particle Identification with the ECL

AD

C c

ou

nts

time

AD

C c

ou

nts

time

8

fast „hadron“ scintillation component


Pulse Shape Discrimination• Multi-template fit (photon and hadron template)

• Offline

• Reconstruct Etotal, Ehadron

Hadron Intensity = EhadEtot

Savino Longo, https://dspace.library.uvic.ca/handle/1828/11301


Keywords to remember:

• Pulse shape:• 31 ADC points

• Difference between photon-like and hadron-like pulse shapes

• Default methods:• FPGA: Etotal, optimized for photons (online)

• Multi-template: Etotal, Ehadron (offline)

• Hadron Intensity = EhadEtot


Neural Network approach

Multi-template fit

Neural Network

Etot, Ehad

What can we gain by using a Neural Network instead of a multi-template fit? (speed, precision, robustness)

?

AD

C c

ou

nts

time


Neural Network structure

• Fully connected NN

• 2 hidden layers (256 neurons each)

• 31 inputs (normalized ADC counts)

• 2 outputs (Etot, Ehad)


Neural Network inputParticle:

Photonsas proxy for EM interactions

Pionsas proxy for hadronic interactions



Stability tests:Baseline

fluctuationsDifferent particle

test data

AD

C c

ou

nts

time







test data

AD

C c

ou

nts

time




Neural Network input photons

• Trained with ~400 000 pulse shapes

• Tested with ~80 000 pulse shapes


Neural Network Etotal prediction

linear response


Neural Network Etotal resolutionEtot resolutiondifference of true value and prediction


energy resolution improved




test data

AD

C c

ou

nts

time




Baseline fluctuations: fittype (FT)

• Assigned by multi-template fit• Fittype = 0 good• Fittype = 1 additional peak• Fittype = -1 bad


Baseline fluctuations: fittype (FT)





test data

AD

C c

ou

nts

time




Neural Network input pions

• Trained with ~328 000 pulse shapes

• Tested with ~82 000 pulse shapes


Neural Network Etot & Ehad resolutionEtot resolution Ehad resolution

hadron resolution improved for photon-like pulse-shapes





test data

AD

C c

ou

nts

time




time

Train with pions, infer on photonsEtot resolution Ehad resolution

total and hadron energy resolution improved up to 200 MeV


What is next?

Running Neural Network

Data & Analysis

• Move towards data• Look at 𝜋𝜋𝛾 analysis• See if NN can improve 𝜋/𝜇-separation• Reduce 𝜇-background• Reduce systematic uncertainty

Belle II, https://docs.belle2.org/record/1124/files/BELLE2-NOTE-PL-2018-032.pdf


Summary

• Running Neural Network for pulse shape analysis

• Performance compared with multi-template and FPGA fit

• Results:• Improved Etot and Ehad-resolution (for pions up to 200 MeV)

• Stable with respect towards fittype

• Interesting application: 𝑒+𝑒− → 𝜋+𝜋−𝛾

• Implement ML on FPGA (new PhD project starting soon)


Back Up.

Traditional Particle Identification

• Measure track → momentum (CDC)

• Measure velocity (TOP)

• Combine information for mass

• 𝜋/𝜇-separation hard, tracks too close


• High precision measurement of the anomalous magnetic momentum of the muon

• 3.7𝜎 discrepancy (arXiv:2006.04822, 2020 paper)

• Theory limited by 𝑒+𝑒− → 𝜋+𝜋−

Excursion: g-2 muon

A. Keshavarzi, D. Nomura, T. Teubner arXiv:1911.00367 [hep-ph]

Theory Experiment


https://arxiv.org/abs/2006.04822

Excursion: theory of g-2

QCD:Needs to be

measured

A. Hoecker and W. J. Marciano, 2013 „The Muon Anomalous Magnetic Moment“


Excursion: theory of g-2

A. Hoecker and W. J. Marciano, 2013 „The Muon Anomalous Magnetic Moment“

QED Universality

Improve precision of g-2 theory by improving precision of the

measurement of this process

T. Aoyama et al. arXiv:2006.04822 [hep-ph]



Analysis: 𝑒+𝑒− → 𝜋+𝜋−𝛾

arXiv:2006.04822 [hep-ph]



FPGA Fit within the ECLA. Sibidanov, https://docs.belle2.org/record/800/files/BELLE2-TALK-CONF-2018-019.pdf

• Two photodiodes• Pre-amplifier (integrate and shape the signal)• Sum• ShaperDSP (shaping amplifier, tail subtraction)• Digitizer (into 31 ADC points)• FPGA (template fit, photon)

Savino Longo, https://dspace.library.uvic.ca/handle/1828/11301 - Chapter 6.1


https://dspace.library.uvic.ca/handle/1828/11301

• All pulse shapes with energy above 50 MeV are stored for offline analysis

• Photon template: • pure photon scintillation component

emission• Use Bhabha scattering events for

computing the templates

• Hadron template• pure hadron scintillation component

emission• Compute with input from signal chain

response and testbeam study

• Need to compute signal chain output for calibration (11 parameters for template for each crystal)

Multi-template fit

Savino Longo, https://dspace.library.uvic.ca/handle/1828/11301 - Chapter 6.3

Savino Longo, https://dspace.library.uvic.ca/handle/1828/11301


https://dspace.library.uvic.ca/handle/1828/11301

Cluster resolution


Neural Network hadron intensity


Beam background Trained with run 3363, infered on run 3363 and 5649 separatly


Neural Network based pulse shape analysis with the Belle ...

Documents

Transcript of Neural Network based pulse shape analysis with the Belle ...